CN110992390A

CN110992390A - Hyperspectral image mixed pixel decomposition method

Info

Publication number: CN110992390A
Application number: CN201911127059.6A
Authority: CN
Inventors: 谭琨; 祝伟; 王雪; 杜培军; 丁建伟
Original assignee: East China Normal University
Current assignee: East China Normal University
Priority date: 2019-11-18
Filing date: 2019-11-18
Publication date: 2020-04-10
Anticipated expiration: 2039-11-18
Also published as: CN110992390B

Abstract

The invention discloses a mixed pixel decomposition method of a hyperspectral image, which is characterized in that a spatial preprocessing and regular hexagon initialization segmentation technology is adopted to segment the image into regions with higher spectral correlation and spatial correlation, PCA projection is carried out on the regions with high correlation, and pixels at positions near the extreme value of a projection axis are selected to select candidate end members. Compared with the prior art, the method has the advantages that the number of pixels participating in end member extraction is greatly reduced, the problem that a large amount of data causes failure of a plurality of mixed pixel decomposition algorithms is effectively solved, and a simple and efficient decomposition algorithm is provided for processing a large amount of data of the hyperspectral remote sensing image.

Description

Hyperspectral image mixed pixel decomposition method

Technical Field

The invention relates to the technical field of remote sensing image processing, in particular to a hyperspectral image mixed pixel decomposition algorithm which is based on space preprocessing and segmentation processing and can be used for processing a large amount of data.

Background

The hyperspectral image contains hundreds of wave bands and contains rich spectral information, so that the spectral characteristics of different ground objects can be completely reflected, and the spectral information provides a large amount of available information for ground object identification. However, due to the limitation of spatial resolution and the complexity of the surface features, mixed pixels are ubiquitous in a hyperspectral image, the mixed pixels seriously influence the identification and interpretation of the surface feature types, and the mixed pixel decomposition is a key technology for solving the problem. On the other hand, because a hyperspectral image contains hundreds of wave bands, the hyperspectral image usually has huge data volume, one band of an aviation hyperspectral image can reach dozens of GB, and in the face of the huge data volume, the prior art usually adopts a blocking processing method, but the method cannot accurately estimate the number of end members of each block, so the unmixing result is restricted, and how to accurately realize the mixed pixel decomposition of the image with huge data volume is a problem to be solved urgently.

Disclosure of Invention

The invention aims to provide a hyperspectral image mixed pixel decomposition method which is designed aiming at the defects of the prior art, the image is divided into areas with higher spectral correlation and spatial correlation by adopting a spatial preprocessing and regular hexagon initialization division technology, PCA projection is carried out on the areas with high correlation, and pixels at positions near the extreme value of a projection axis are selected to select candidate end members.

The specific technical scheme for realizing the purpose of the invention is as follows: a mixed pixel decomposition method of a hyperspectral image is characterized in that a spatial preprocessing and regular hexagon initialization segmentation technology is adopted to segment an image into regions with high spectral correlation and spatial correlation, PCA projection is carried out on the regions with high correlation, and pixels near the extreme value of a projection axis are selected to select candidate end members, and the specific method comprises the following steps:

step 1: and acquiring hyperspectral image data, and estimating the number of end members of the hyperspectral image.

Step 2: the method comprises the steps of initializing by adopting a honeycomb form, segmenting an image, dividing an original image into a plurality of hexagons according to the set average size h of each hexagon, determining the number of initialized segmentation blocks and the center of each block according to the geometric properties of the hexagons, endowing different labels to pixels of each block, and defining the initial distance from the center of each block to all pixels in the block to be infinite.

And step 3: in each block, the Spectral Distance between the center pixel and its surrounding pixels of each block is calculated, where the Spectral Distance can be a variety of Spectral Distance measurement criteria, such as Spectral Correlation Angle (SCA), Spectral Information measure (SID), Spectral Angle Distance (SAD), or their combined Spectral Information Divergence-Spectral Correlation coefficient (SID-SCA).

And 4, step 4: in each block, the spatial euclidean distance of the central pixel of each block to its surrounding pixels is calculated.

And 5: combining the step 3 and the step 4 to obtain a joint distance m, and judging the distance between the central pixel and the peripheral pixels of each block according to the distance, wherein the joint distance m is calculated according to the following formula a:

in the formula: c is the spectral distance, s is the Euclidean distance, the values of the Euclidean distances are respectively obtained in the step 3 and the step 4, l is the length of the diagonal line of the hexagon, a is the combined weight, and 0< a < 1.

Step 6: updating each block and judging: if the joint distance m from the center of each block to any pixel is less than its previous value (the initial value of this distance has been defined to be infinite in step 2), its distance and label are updated.

And 7: update the center pixel of each block: the average spectral information for each block is calculated, with this value as the new center pixel spectral information.

And 8: and (5) repeating the steps 3 to 7 until the preset repetition number Iter is reached.

And step 9: and removing the isolated small region, setting the side length as x, and combining the region with the side length smaller than x into adjacent pixels to obtain final block information.

Step 10: and carrying out PCA projection on each block, selecting the first q principal component vectors as projection axes, projecting all pixels in the block onto each projection axis, and recording the projection positions.

Step 11: selecting pixels at the positions of two ends of the projection as target end member signals, recording the maximum value as max and the minimum value as min, and calculating the projection values pri of the other pixel points according to the following formula b:

in the formula: and p is a projection value of each pixel on the projection axis, the maximum value of the projection value is max, and the minimum value of the projection value is min.

Step 12: calculating the projection weight wi of each pixel according to the following formula c:

step 13: according to the contribution ratio cj of each pixel on the q projection axes, calculating the weighted sum of the projection values of each pixel on each projection axis according to the following formula d as a spectral index spi:

in the formula: n is the number of pixels in the image; q is the number of projection axes.

Then, the spectral indexes spi are sorted, and according to the pixel percentage pe in each block, the pixel located at the top pe in spi is selected as a candidate pixel.

Step 14: end-members are extracted from the candidate pixels using a spectral-based end-member extraction algorithm.

Compared with the prior art, the method has the advantages that the image mixed pixels with huge data volume are quickly decomposed, the problem that a plurality of mixed pixel decomposition algorithms are invalid due to the huge data volume is effectively solved, a simple, convenient and efficient decomposition algorithm of the image mixed pixels is provided for processing the hyperspectral remote sensing images with large data volume, and the wide application of the hyperspectral remote sensing images is further promoted.

Drawings

FIG. 1 is a flow chart of the present invention;

FIG. 2 is a shape of a segmentation initialization;

FIG. 3 is a graph of a segmentation result of the synthesized hyperspectral image data;

FIG. 4 is an end-member graph extracted from the synthesized hyperspectral image data.

Detailed Description

The present invention will be described in further detail below with reference to the synthesized hyperspectral image as an example.

Example 1

Referring to the attached figure 1, the method processes the mixed pixel decomposition of the hyperspectral image with large data volume according to the following steps:

acquiring hyperspectral image data

The synthesized hyperspectral image data is obtained, the size of the hyperspectral image data is 90 multiplied by 90, the number of wave bands is 224, and the number of end members of the image is 5.

(II) image segmentation

Referring to fig. 2, an original image is divided into a plurality of hexagons by adopting honeycomb-shaped initialization, the average size h of each hexagon is set to be 7, the number of initialized divided blocks and the center of each block are determined according to the geometric properties of the hexagons, different labels are given to pixels of each block, and the initial distance from the center of each block to all pixels in the block is defined to be infinite.

(III) calculation of spectral distance c

In each block, the spectral information divergence-spectral correlation coefficient (SID-SCA) of the central pixel of each block and its surrounding pixels is calculated as the spectral distance c.

(IV) calculation of the spatial Euclidean distance s

In each block, the spatial euclidean distance s of the central pixel of each block and its surrounding pixels is calculated.

(V) calculation of the Joint distance m

Calculating the joint distance m according to the following a formula, and judging the distance between the central pixel and the surrounding pixels of each block according to the distance:

in the formula: c is a spectral distance obtained in the step (three); s is the Euclidean distance, and is obtained by the step (IV); l is the length of the diagonal of the hexagon; a is the combined weight, 0< a <1, where a is 0.1.

(VI) updating each block

And (3) judging: if the combined distance m from the center of each block to any pixel is less than its previous value, its distance and label are updated.

(VII) updating the center pixel of each block

The average spectral information for each block is calculated, with this value as the new center pixel spectral information.

(eight) cycle update

Repeating the steps (three) to (seven) until a preset repetition number Iter is reached, wherein Iter of the embodiment is set to 10.

(nine) removing isolated small regions

Referring to fig. 3, the side length x is set to be 2, and for small regions with side lengths smaller than x, the small regions are combined into adjacent pixels to obtain final block information.

(ten) PCA projection

And carrying out PCA projection on each block, selecting the first q principal component vectors as projection axes, projecting all pixels in the block onto each projection axis, and recording the projection position, wherein q is equal to 3.

Calculation of the (eleven) projection value pri

Selecting pixels at the positions of two ends of the projection as target end member signals, recording the maximum value as max and the minimum value as min, and calculating the projection values pri of the other pixel points according to the following formula b:

in the formula: p is a projection value of each pixel on the projection axis.

(twelve) calculation of projection weight wi

Calculating the projection weight wi of each pixel according to the following formula c:

(thirteen) calculation of the spectral index spi

According to the contribution ratio cj of each pixel on the q projection axes, calculating the weighted sum of the projection values of each pixel on each projection axis as a spectral index spi, wherein the calculation formula is as follows:

Then, spi is sorted, and pixels located at the top pe of 10% in spi are selected as candidate pixels according to the defined percentage pe in each block.

(fourteen) terminal member extraction

Referring to FIG. 4, a spectral-based end-member extraction algorithm uses Vertex Composition Analysis (VCA) to extract end-members from candidate pixels.

The invention has been described in further detail in order to avoid limiting the scope of the invention, and it is intended that all such equivalent embodiments be included within the scope of the following claims.

Claims

1. A mixed pixel decomposition method of a hyperspectral image is characterized in that a spatial preprocessing and regular hexagon initialization segmentation technology is adopted, the image is segmented into regions with high spectral correlation and spatial correlation, PCA projection is carried out on the regions with high correlation, and pixels at positions near an extreme value of a projection axis are selected to select candidate end members, and the specific method comprises the following steps:

step 1: acquiring hyperspectral image data, and estimating the number of end members of a hyperspectral image;

step 2: adopting honeycomb form initialization to segment the image;

and step 3: calculating the spectral distance c between the central pixel of each block and the surrounding pixels;

and 4, step 4: calculating the space Euclidean distance s between the central pixel of each block and the surrounding pixels;

and 5: combining the step 3 and the step 4 to obtain a combined distance m, and judging the distance between the central pixel and the peripheral pixels of each block according to the distance;

step 6: updating each block and judging: if the distance from the center of each block to any pixel is less than its previous value, updating its distance and label;

and 7: update the center pixel of each block: calculating the average spectral information of each block, and taking the value as new central pixel spectral information;

and 8: repeating the steps 3 to 7 until a preset repetition number Iter is reached;

and step 9: clearing isolated small regions: setting the side length as x, and combining the areas with the side length smaller than x into adjacent pixels to obtain final block information;

step 10: carrying out PCA projection on each block, selecting the first q principal component vectors as projection axes, projecting all pixels in the block to each projection axis, and recording the projection position;

step 11: selecting pixels at the positions of two ends of the projection as target end-member signals, recording the maximum value and the minimum value as max and min respectively, and calculating the projection values pri of the rest pixel points;

step 12: calculating a projection weight wi of each pixel;

step 13: calculating the spectral index of each pixel according to the projection weight wi of each pixel, and defining the pixel percentage pe to be selected by each partition to obtain a candidate pixel;

2. The method for decomposing the mixed pixels of the hyperspectral image according to claim 1, wherein the honeycomb form initialization segmentation is to divide the original image into a plurality of hexagonal blocks according to the set average size h of each hexagon, and then to determine the number of initialization segments and the center of each block according to the geometric properties of the hexagons.

3. The method for decomposing the mixed pixels of the hyperspectral image according to claim 1, wherein the spectral distance c is a spectral correlation coefficient, a spectral information measure, a spectral angular distance, or a spectral information divergence-spectral correlation coefficient of a combination of two of the spectral correlation coefficient and the spectral information measure.

4. The method for decomposing the mixed pixels of the hyperspectral image according to claim 1, wherein the joint distance m is calculated according to the following formula a:

in the formula: c is the spectral distance; s is the Euclidean distance; l is the length of the diagonal of the hexagon; a is the combined weight, 0< a < 1.

5. The method for decomposing the mixed pixels of the hyperspectral image according to claim 1, wherein the projection values pri of the remaining pixels are calculated according to the following formula b:

6. The method for decomposing the mixed pixels of the hyperspectral image according to claim 1, wherein the projection weight wi of each pixel is calculated according to the following formula c:

7. the method for decomposing the mixed pixels of the hyperspectral image according to claim 1, wherein the spectral index spi is a contribution ratio cj of each pixel on q projection axes according to the following formula d, and a weighted sum of projection values of each pixel on each projection axis is calculated according to the following formula d to serve as the spectral index spi:

in the formula: n is the number of pixels in the image; q is the number of projection axes; cj is the contribution ratio of each of the q projection axes.

8. The method for mixed pixel decomposition of hyperspectral image according to claim 1, wherein the candidate pixels are spectral indexes spi ranked, and then according to the pixel percentage pe in each block, the pixel at pe before spi in spi is selected as the candidate pixel.